Tracking changes in behavioural dynamics using prediction error
Fig 3
Prediction error reveals anomalous dynamics in worm pose time series.
A) Original time series of eigenworm coefficients indicating the split into reference library and prediction set. Note that values in the prediction set are only predicted one time step ahead based on the previous embedded value from the prediction set; i.e., we are not “forecasting” 60 seconds into the future (see Method Overview and Methods and materials). Error peaks in (B) indicated by arrows do not correspond to obvious features in the time series. B) Prediction error corresponding to prediction set in (A). Turns (red dots), characterised here by third eigenworm coefficient |a3|>15, are book-ended by periods of anomalous dynamics. The mean error for a constant predictor (i.e. the prediction that the worm pose is the same as the previous time step) is higher than the vertical axis range (0.126). C) A closer look at the worm pose time series (black) and predictions (orange) reveals that the anomalous dynamics correspond to delta turn initiation and completion, when the worm head and tail respectively transition between self-intersection and no self-intersection.